IRJET- Intelligent Depression Detection System

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 03 | Mar 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 2934
INTELLIGENT DEPRESSION DETECTION SYSTEM
B.Yamini 1, S.Aruljothi 2, S.M Madhumitha3
1Assistant Professor, Computer Science Department, JEPPIAAR SRR Engineering College, Tamil Nadu
2,3UG Student, Computer Science Department, JEPPIAAR SRR Engineering College, Tamil Nadu
----------------------------------------------------------------------***---------------------------------------------------------------------
Abstract-- Depressionis becoming a serious leading mental health problem worldwide. It is a cause ofpsychologicaldisordersandeconomic
burden to a country. Even though it is a serious psychological problem,lessthanahalfofthosewhohavethisemotionalproblemgainedaccess
to mentalhealth service. This could be a result ofmany factors includinghavinglackawarenessaboutthedisease.Newsolutionsareneededto
tackle this issue. We propose a system is to developprediction models to classify users according to depressionlevels fromFacebookdataby
employingNatural Language Processing (NLP) techniques where people use Facebook as a toolfor sharing opinions, feelingsandlifeevents.
Keywords -- Natural Language Processing, Statistical analysis, Facebook site, Depression screening, User Generated Content (UGC).
1. Introduction
Social mediacanbeexploitedduetothelarge amountofinformation,whichreferstouserbehavioralattributes.Gettinguseof
that information to predict the social media users’ mental health level can help psychiatrist, family or friends to get the right
medicaladvice and therapy ontimeto the depressed user.WHOranks the depression asone ofthe mostdevastatingdiseasesin
the world. In addition, about two thirds of depressed people do not seek appropriate treatments, which lead to major
consequences. The medical science relies on asking the patientsquestions about their circumstances, which does not diagnose
the depression in a precise way. The patient has to attend more than one session during a period of three weeks. The
classification of a not depressed condition as a depressed is a False Positive problem. However, researchers found that the
Electronic Health Record (EHR) systems are not optimally designed to handle integrating behavioral health and primary care.
EHRslack to support documenting and trackingdataforbehaviouralhealthconditionssuchasdepression.Mostofthepeopleuse
social mediato express their feelings, emotions. Many researchers have been successfully proving that social media has been
successfully used to maintainpeople's mental health. By mining the social media posts ofusers, wemaygeta completeimageof
theusernatural behaviour, thinking style,interactions,guiltfeeling,worthlessness,loneliness,andhelplessness.Retrievingsuch
behavioural attributes, show symptoms of depression on the social media users, which could be used to predict if the user is
depressed or not. Psychiatrist, parents, and friends, could track the user depression by the proposed tool, which will save the
time before the depressed user could get into major depression phase.
2. Related Work
Several approaches have been studied for collecting social media data with associated information about the users’ mental
health. Participants are either recruited to take a depression surveyandsharetheirFacebookorTwitterdata,ordataiscollected
from existing public online sources. These sources include searching public Tweets for keywords to identify (and obtain all
Tweets from) users who have shared their mental health diagnosis, user language onmental illness related forums, orthrough
collecting public Tweets that mention mental illness keywords for annotation. The approaches using public data have the
advantage that much larger samples can, in principle, be collected faster andmore cheaply than through the administration of
surveys, though survey-based assessment generally providesahigherdegreeofvalidity.Wefirstcomparestudiesthatattemptto
distinguish mentally ill users from neurotypical controls.
A. Prediction Based on Survey Responses
Psychometric self-report surveys for mental illness have a high degree of validity and reliability. In psychological and
epidemiological research, self- reportsurveysaresecondonlytoclinicalinterviews,whichnosocialmediastudytodatehasused
as an outcome measure. We discuss five studies that predict survey assessed depression status by collecting participants’
responses to depression surveys in conjunction with their social media data. The most cited study used Twitter activity to
examine network and language data preceding a recent episode of depression. The presence of depression was established
through participants reporting the occurrence and recent date of a depressive episode, combined with scores on the Center
forEpidemiologic Studies Depression Scale Revised(CES-D) and Beck’s DepressionInventory(BDI).Thisstudyrevealedseveral
distinctions in posting activity by depressed users,including:diurnalcycles,morenegativeemotion,lesssocialinteraction,more
self- focus, and mentioning depression-related termsthroughouttheyearprecedingdepressiononset.Reeceetal.predicteduser
depression and posttraumatic stress-disorder (PTSD) status from text and Twitter meta-data that preceded a reported first

episode with relatively high Areas under the Receiver Operating Characteristic (ROC) curve (AUCs) of .87 (depression) and.89
(PTSD). Data were aggregated to weeks, which somewhat outperformed aggregation to days, and modeled as longitudinal
trajectories of activity patterns that differentiated healthy from mentally ill users. Tsugawa et al predicted depression from
Twitter data in a Japanese sample, using the CES-D as theirassessment criterion. Using tweetsfromthemostrecent6–16weeks
preceding the administration of the CES-D was sufficient for recognizing depression; predictions derived from data across a
longer period were less accurate. While most studies have used Twitter, used Facebook status updates for the prediction.
Mothers self-reported a specific postpartum depression (PPD) episode and completed a screening survey. A model using
demographics, Facebook activity, and content of posts before childbirth accounted for 35.5% of the variance in PPD status.
Schwartz et al used questionsfromapersonalitysurveyto determineusers’continuousdepressionscoresacrossalargersample
of Facebook users (N = 28 749) than used in other studies (which typically range in the low hundreds). This study observed
seasonal fluctuations of depression, finding that people were more depressed during winter months.Thisstudyalsoprovideda
shortlist of the words, phrases and topics (clusters of semantically coherent words) most associated with depression. Survey
responses provide the most reliable ground- truth data for predictive models in this emerging literature.
B. Prediction Based on Self-declared Mental Health Status
A number of studies use publicly accessible data. ‘Self-declared’ mental illness diagnosis on Twitter (identified through
statements such as ‘I was diagnosed with depression today’) is one such source of publicly-available data. We review seven
studiesof this kind. Helping to facilitate studies of this kind, a Computational Linguistics and Clinical Psychology (CLPsych)
workshop was started in 2014 to fostercooperation betweenclinical psychologistsandcomputerscientists.‘Sharedtasks’were
designed to explore and compare different solutions to the same prediction problem onthe same data set.Inthe 2015 CLPsych
workshop, participants were asked to predict if a user had PTSD or depression based on self- declared diagnoses on Twitter
(PTSD n = 246, depression n = 327, with the same number of age- matched control and gender-matched control). Participating
teams built language topic models (e.g. an anxiety topic contained the words: feel, worry, stress, study, time, hard), sought to
identify words most associated with PTSD and depression status, considered sequencesof characters as features,andapplieda
rule-based approach to buildrelative counts ofN-grams present in PTSD and depression statusesofallusers.Thelatterresulted
in the highest prediction performance. All approaches found that it was harder to distinguish between PTSD and depression
versus detecting the presence of either condition (compared to controls), suggesting overlap in the language associated with
both conditions. On a shared dataset similar to the 2015 CLPsych workshop, the prediction of anxiety was improved by taking
genderintoaccount in additionto10comorbidconditions.Otherstudieshaveusedpsychologicaldictionaries(LinguisticInquiry
and Word Count; LIWC) to characterize differences between mental illness conditions, withsomesuccess. Onthesamedataset,
Preotic-Pietroetal observedthatestimatingtheageofusers adequatelyidentifieduserswhohadself-declaredaPTSDdiagnosis,
and that the language predictive of depression and PTSD had large overlap with the language predictive of personality. This
suggests that users withparticular personality ordemographic profileschosetosharetheirmentalhealthdiagnosisonTwitter,
and thus that the results ofthese studies (mostly, predictionaccuracies)maynotgeneralizetoothersourcesofautobiographical
text.
C. Prediction Based on Forum membership
Online forums and discussion websites are a secondsource ofpublicly-availabletextrelatedtomentalhealth.Theyofferaspace
in which users can ask for advice, receive and provide emotional support, and generally discuss stigmatized mental health
problems openly. We review three such studies here. Theforum (reddit)posts wereused tostudythemental well-being of U.S.
university students. A prediction model was trained on data gathered from reddit mental health support communities and
applied to the posts collected from 109 university subreddits to estimatethe level of distress at theuniversities.Theproportion
of mental health posts increased over the course of the academic year, particularly for universities with quarter-based, rather
than semester-based, schedules. The language of 16 subreddits covering a range of mental health problems was characterized
usingLIWC and other markers ofsentence complexity. DeChoudhury et al examinedpostsofagroupofreddituserswhoposted
about mental health concerns and then shifted to discuss suicidal ideation in the future. Several features predicted this shift:
heightened self- focus, poor linguistic style matching with the community, reduced social engagement, and expressions of
hopelessness, anxiety, impulsiveness, and loneliness.
D. Prediction based on Annotated Posts
The publicly-available text involves manually examining and annotating Tweets that contain mental health keywords.
Annotator’s code social media posts according to pre-established (a priori, theory driven) or bottom-up (determined from the
data) classifications; annotations can be predicted from the language of posts. 46 Big data in the behavioural sciences 1 that is,

not using cross- validation. Most annotation studies on depression focus on identifying postsinwhichusersarediscussingtheir
own experience with depression. Annotators are provided with guidelines on how to recognize a broad range of symptoms of
depression thatarederivedfromclinicalassessmentmanualssuchastheDSM-5,orareducedsetofsymptoms,suchasdepressed
mood, disturbed sleep and fatigue .Annotation has also been used to differentiate between mentions of mental illness for the
purpose of stigmatization or insult as opposed to voicing support or sharing useful information with those suffering from a
mental illness. In general, annotations of posts are a complementary (but labor-intensive) method that can reveal life
circumstances associated with mental illness not captured by traditional depression diagnostic criteria.
3. Methodology
The Depression prediction model uses Natural Language Processing. The model consists of two datasets, and seven main
operators. The first dataset isthetrainingdatasetwhichcontainsthemanuallytraineddepressedpostsandnot-depressedposts.
The second dataset consists of the patient Facebook posts and it is changed for every individual to test the prediction of the
model.Thefirstoperatoris the Select Attributes which selects which attributes of thetrainingdatasetshouldbekeptandwhich
attributes should be removed. The second and the third operators are the Nominal to Text, this operator changes the type of
selected nominal attributes to text, also it maps all values of the attributes to corresponding string values, it is used in the
training dataset and the test set. The fourth and the fifth process are Process Process Documents and it is used in the training
datasetand the test set whichgenerates word vectors fromstring attributesand itconsists offouroperators.Thefouroperators
of Process Document operator are Tokenize, Filter Stop- words, Transform Cases, and Stem. The Tokenize operators splits the
text of a document into a sequence of tokens. The filter Stop-words filters English stop words from a document by removing
every token which equals a stop-word from the built-in stop-word list. The Transform Cases operatortransformsallcharacters
in a document to lowercase. The Stem operator stems English words using the Porter stemming algorithm intending to reduce
the length of the words until a minimum length isreached.The sixth operator is the Validation operator which contains which
applies onthetrainingdataset which consistsoftwo main sections, training,and testing. The seventh andlast operator is Apply
model which connect the testdataset and training dataset togiveusthefinalresultofthepredictionusingoneoftheclassifiersin
the patients. The accuracy of the classification depends on the training set used to run the Natural Language Processing.
4. Proposed System
The approach being taken up in this paper is modular in its’ organization. Individual component of the work flow has been
segregated into stand-alone steps, toimprove qualityofimplementation. The work flow starts with data collection step, which
utilizes Facebook API fora generation ofdataset.Naturallanguage processing facilitatesmuchbetterthanaverageclassification
for sentimental analysis done by the human. Following the creation of datasets, the data pre processing module which
systematically churns the data through tokenization, stemming and stop words removal. POS tagger then identifies essential
pieces of the text to be utilized. After this the text classifier is trained on the processed text data from Facebook, in the training
phase. In the testing phase, class prediction is made on the test dataset to identify potential Text demonstrating depression
tendencies.
Figure 1 Data Collection

A) Data Collection
Sentiment analysis task starts with a collection of relevantdatafromvarioussources.InthisFacebookisconsideredasthedata
source for analysis, in the form of User Generated Content. This portion covers tasks from streaming the data, to compile the
training and test datasets.
1) Facebook API: By Facebook API, we collect the user generated content specifically images.
2) Keyword List: Using a pre-created wordlist fordetectingtriggerwordssymbolizingpoormentalwell-being, Textfromall over
the world are collected at random. These keywords specific texts are mixed with a general batch of non-weighted Texts.
3) Extracting Text from JSON: The collected text in the JSON objects are parsed to extract only the text field of the Text. Other
meta-data related to any particular Text is removed.
4) Data Cleaning: To avoid errors in encoding textual data, the Text is purged for links (http) and non-ASCII characters like
emoticons. Result is a clean dataset, rid of non-confirmative character types.
5) GenerateCSVFileforTrainandTestSet: Thecleaned text data from individual Text is added to the training and test dataset, in a
vectorized format. Classification labels for the training and test datasets are manually added,tocreatea CSVfileusing comma
as the delimiter.
B) Data Pre-processing:
The CSV file is read and several data pre-processing steps are performed on it. Natural language processing has been utilized
for pre-processing methods applied on the extracted data:
1) Tokenization: Tokenization is a process of dividing a string into several meaningful substring, such as units of words,
sentences, or themes. In this case, the first column of the CSV file containing the tweet is extracted and is converted into
individual tokens.
2) Stemming: Stemming involves reducing the words to their root form. This would help us to group similar words together.
For implementation, Porter Stemmer is used.
3) Stop Words Removal: The commonly used words, known as stop wordsneed toremove since they are of no use in the
training and could also lead to erratic results if not ignored. Nltk library has a set of stop words which can be used as a
reference to remove stop words from the text.
4) POS Tagger: To improve the quality of the training data, the tokenized text is assigned the respective parts of speech by
using POS Tagger. This wouldbeusedtoextractonlytheadjectives,nounsandadverbssinceotherpartsofthespeecharenotof
much significance.
After all these pre-processing steps, a bag of words is formed. Bag of words calculates the number of occurrences of each
word, which is then used as a feature to train a classifier.
C) Training
The classifier requires two parameters: training set and label. The training set in this case is the set of tweets which needs to be
further processed in ordertofeedintoa classifier.Theset oftweetsneedtoconvertintovectorformatforfurtherprocessing.The
set of labels corresponding to each text is alsofed into the classifier in the form a vector. Saving the Classifier and the Count
Vectorized Object: Since training needs to be done once, the trained classifier object needs to be loaded intoapicklefile.Sameis
applicable with the Count Victimizer object. Thus, both these objects are dumped into a pickle file for further use.

Fig. 2 Training Phase
Figure 3 Testing Phase
D) Testing
1) Loading: Classification models are loaded from the pickle file, to be used for prediction on test dataset.
2) Data Preprocessing: The test dataset is preprocessed in a manner similar to the training data.
3)Class Prediction on Test Text: Each tweet is classified into a depressed or not depressed or neutral class.
4) Comparison Based onthe class prediction: At lastthecomparetheallclasspredictionsetandfinalizetheresultwhether
the person in depressed stateornot. Further depression state is classifiedintothreestate(MINIMAL,MAXIMALAND
SEVERE) and also, we give a suggestion according to their person depression state and nearby doctors location.
5) Result
To analyze the performance ofour proposed model, we calculated precision and recall. The precision gives us the percent ofthepositive
correct prediction of processed images and recall give us the percent of the false prediction of processed images. The performance was
tested with up to 100 images. Table 1 shows computing metrics of SVM classifier. Table 2 shows the value for each True positive, True

Negative, False Positive, False Negative of SVM classifier in depression prediction.
TABLE 1
Computing Metrics
TABLE 2
Comparative Analysis Of Related Work
Fig 3. A Sample Bar Graph Using Precision and Recall Value.
Evaluating the system comparing with related work in Table 6 wherewe compare the performance metricsoftheproposedsystemandthe
related works. Our proposed system got the best precision and least in recall.
5. Discussion
The greatest potential value ofsocial media analysis may bethe detection ofotherwise undiagnosed cases. However, studies to
date have not explicitly focused on successfully identifying people unaware of their mental health status. In screening for
depression, multi-stage screening strategies have been recommended as a means to alleviate the relatively low sensitivity
(around 50%) and high false positive rate associated with assessments by non-psychiatric physicians or short screening
inventories. Social- media based screening may eventually provide an additional step in a mental health screening strategy.
Studies are needed that integrate social media data collection with gold standard structured clinical interviews and other
Metrics Text in
Images
Classifier
Output
True
Positive(TP)
Positive data Positive Data
True
Negative(TN)
Negative Data Negative Data
False
Positive(FP)
Negative Data Positive Data
False
Negative(FN)
Positive Data Negative Data
Total
Image
TP TN FP FN Precisi
on
Recall
10 5 3 2 1 0.714 0.833
30 10 12 5 3 0.666 0.769
50 21 11 9 9 0.7 0.7
70 30 25 11 4 0.731 0.882
100 45 39 8 8 0.849 0.849

screening strategies in ecologically valid samples to test the incremental benefit of social media-based screening and
distinguishing betweenmental healthconditions.Self-reportedsurveysandclinicaldiagnosesprovidesnapshotsintime.Online
social media data may ‘fill in the gaps’ withongoing in-the-moment measures ofa broad range ofpeople’sthoughtsandfeelings.
However, as depressed users may cease generating social media content, alternative uninterrupted data streams such as text
messages and sensor data should also be tested for ongoing monitoring applications.
7. Conclusion:
Psychological depression is threatening people’s health. It is non-trivial to detect depress level timely for proactive care.
Therefore, wepresentedaframeworkfordetectingusers’psychologicaldepressionstatesfromusers’monthlysocialmediadata,
leveraging face book post’ content as well as users’socialinteractions.Employingreal-worldsocial media data as the basis, the
project goal is to develop a web application which takes social media posts and predict output as various depression levels
(minimal, maximaland severe).Using Natural Language Processing algorithm toincreasetheaccuracyofsystem.Anditsdeliver
appropriate doctor's information depending upon location of user. According to his Facebook post system can find out user in
depressed or not which is provided by the system.
REFERENCES
[1] S. R.Chowdhury,M. Imran, M. R.Asghar, S.Amer- Yahia, and C.Castillo, “Tweet4act: Using incident-specificprofilesforclassifyingcrisis-
related messages,” in 10th Proc. Int. Conf. Inform. Syst. Crisis ResponseManag., 2013.
[2] T. Davidson, D. Warmsley , M.W.Macy, and I.Weber,“Automated hate speech detection and the problem of offensivelanguage,”
inProc.11thInt.Conf.WebSocialMedia,2017,pp.512–515 T. Hofmann, “Probabilistic latentsemantic indexing ,” inProc.An
nu.Int.ACMSIGIRConf.Res.Develop.Inf.Retrieval,1999,pp. 50–57.
[3] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent dirichlet allocation,” J. Mach. Learn., vol. 3, pp. 993–1022, 2003.
[4] Y. Wang, E. Agichtein, and M. Benzi, “TM-LDA: Efficient online modeling of latent topic transitions in social media,” in Proc. ACM
SIGKDDInt.Conf.Knowl.Discovery DataMining, 2012, pp.123–131.
[5] S.Sidana,S.Mishra,S. Amer-Yahia, M. Clausel,andM. Amini, “Health monitoring onsocial media overtime,” in Proc. 39th Int. ACM SIGIR
Conf. Res. Develop. Inf. Retrieval, 2016, pp. 849–852.
[6] K. W.Prier, M. S. Smith,C.Giraud-Carrier, and C.L.Hanson,“Identifyinghealth-relatedtopicsontwitter,”inSocialComputing,Behavioral-
Cultural Modeling and Prediction. Berlin, Germany: Springer, 2011, pp. 18–25.
[7] C. Cortes and V. Vapnik, “Support-vector networks,” Mach.Learn., vol. 20, no. 3, pp. 273–297, 1995. [Online].
Available: http://dx.doi.org/10.1007/BF00994018
[8] M. De Choudhury, “Anorexia on tumblr: A characterization study,” in Proc. 5th Int. Conf. Digital Health, 2015, pp. 43–50.
[9] M. De Choudhury,A.Monroy-Hern andez, and G. Mark, ““narco” Emotions: Affectanddesensitizationinsocialmediaduringthemexican
drug war,” inProc. SIGCHI Conf. Human Factors Comput.Syst., 2014, pp. 3563–3572.
[10] U. Pavalanathan and M. De Choudhury,“Identity management and mentalhealth discourse in the social media,” in Proc.24thInt.Conf.
World Wide Web, 2015, pp. 315–321.
[11] F.Bouillot,P.Poncelet,M.Roche,D.Ienco,E.Bigdeli, and S.Matwin,“Frenchpresidentialelections:Whatarethe mosteffi-cientmeasures
for tweets?” in Proc. 1st Edition Workshop Politics Elections Data, 2012, pp. 23–30.
[12] L. Hemphill and A. J. Roback, “Tweet acts: How constituents lobby congress via twitter,” in Proc. ACM Conf.Comput. Supported
Cooperative Work Social Comput., 2014, pp. 1200–1210.
[13] A.Ceron, L.Curini, and S. M. Iacus,“Using sentiment analysis to monitorelectoralcampaigns:Methodmatters-evidencefromthe United
States and italy,” Social Sci. Comput. Rev., vol. 33, no. 1 pp. 3–20, 2015.
[14] P. Barbera, “Birds of the same feather tweet together: Bayesian idealpointestimationusingtwitterData,”Political Anal., vol.23, no. 1,
pp. 76–91, 2015.
[15] L. Jiang, M. Yu, M. Zhou, X.Liu, and T.Zhao, “Target- dependent twitter sentiment classification,” in Proc. Annu. Meeting Assoc.Comput.
Linguistics: Human Language Technol., 2011, pp. 151–160.
[16] B. J. Jansen, M. Zhang, K. Sobel, and A.Chowdury, “Twitter power: Tweets asan electronic word ofmouth,”J.Amer.Soc.Inf.Sci.Technol.,
vol. 60, no. 11,pp. 2169–2188, 2009.
[17] S. Wang, M. J. Paul, and M. Dredze, “Exploringhealth topics in Chinese social media: Ananalysis ofsina weibo,” in Proc.AAAIWorkshop
World Wide Web Public Health Intell., 2014.
[18] A.Culotta, “Estimating county health statistics with twitter,” in Proc.SIGCHIConf. Human Factors Comput.Syst.,2014,pp.1335–1344.

IRJET- Intelligent Depression Detection System

Recommended

Recommended

More Related Content

What's hot

What's hot (19)

Similar to IRJET- Intelligent Depression Detection System

Similar to IRJET- Intelligent Depression Detection System (20)

More from IRJET Journal

More from IRJET Journal (20)

Recently uploaded

Recently uploaded (20)

IRJET- Intelligent Depression Detection System